# Time Series Forecasting Model Leaderboard & Comparisons

> Empirical probabilistic forecasting benchmark evaluating zero-shot foundation models across 100 fev-bench standardized tasks.

## Global FEV Benchmark Leaderboard

| Rank | Model / Method | Family | Organization | SQL Win Rate (%) | SQL Skill Score | MASE Win Rate (%) | Median Latency | Execution |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| 1 | **Timestral-209M** | Foundation Model | — | 87.7% | 50.39 | 86.0% | 0.51s | Server (CUDA / PyTorch) |
| 2 | **Timestral-123M (Carla Engine)** | Foundation Model | — | 86.8% | 50.18 | 85.0% | 0.32s | WebGPU (Browser / Sheets) |
| 3 | **TimeRSI** | Closed-Source API | — | 84.8% | 51.24 | 83.8% | 2512s | Server (CUDA / PyTorch) |
| 4 | **t0-gamma** | Foundation Model | — | 84.4% | 49.42 | 81.8% | 4.38s | Server (CUDA / PyTorch) |
| 5 | **Timestral-28M** | Foundation Model | — | 79.2% | 48.42 | 77.0% | 0.11s | Server (CUDA / PyTorch) |
| 6 | **timesfm-3** | Foundation Model | — | 78.5% | 48.99 | 77.0% | 3.41s | Server (CUDA / PyTorch) |
| 7 | **ChakraTS** | Baseline | — | 78.2% | 48.14 | 79.5% | 14.22s | Server (CUDA / PyTorch) |
| 8 | **Lingjiang2.0** | Baseline | — | 73.9% | 47.02 | 71.3% | 8.89s | Server (CUDA / PyTorch) |
| 9 | **Chronos-2** | Foundation Model | — | 70.4% | 47.28 | 68.4% | 0.84s | Server (CUDA / PyTorch) |
| 10 | **timeh2o-v1-tiny** | Foundation Model | — | 69.9% | 45.49 | 67.0% | 0.19s | Server (CUDA / PyTorch) |
| 11 | **Timestral-5M** | Foundation Model | — | 69.0% | 46.19 | 67.7% | 0.04s | Server (CUDA / PyTorch) |
| 12 | **Toto-2.0-2.5B** | Baseline | — | 67.5% | 44.42 | 65.7% | 5.58s | Server (CUDA / PyTorch) |
| 13 | **t0-beta** | Foundation Model | — | 67.4% | 46.65 | 64.5% | 2.29s | Server (CUDA / PyTorch) |
| 14 | **Toto-2.0-1B** | Baseline | — | 66.4% | 44.43 | 65.5% | 2.48s | Server (CUDA / PyTorch) |
| 15 | **TiRex-2** | Foundation Model | — | 65.8% | 45.50 | 63.4% | 0.27s | Server (CUDA / PyTorch) |
| 16 | **Toto-2.0-313m** | Baseline | — | 65.1% | 44.23 | 63.5% | 1.06s | Server (CUDA / PyTorch) |
| 17 | **TiRex** | Foundation Model | — | 58.0% | 42.58 | 54.1% | 0.24s | Server (CUDA / PyTorch) |
| 18 | **Toto-2.0-22m** | Baseline | — | 57.6% | 42.91 | 56.0% | 0.25s | Server (CUDA / PyTorch) |
| 19 | **TimesFM-2.5** | Baseline | — | 54.6% | 42.20 | 53.5% | 1.87s | Server (CUDA / PyTorch) |
| 20 | **TabPFN-TS-3** | Foundation Model | — | 52.4% | 43.15 | 51.3% | 235s | Server (CUDA / PyTorch) |

## Key Pairwise Comparisons

- [Timestral-123M vs Google TimesFM 3](/datasets/comparison/time-series/timestral-123m-vs-timesfm-3.md): Head-to-head probabilistic loss and latency comparison.
- [Timestral-123M vs AWS Chronos-2](/datasets/comparison/time-series/timestral-123m-vs-chronos-2.md): Head-to-head probabilistic loss and latency comparison.
- [Timestral-123M vs Prior Labs TabPFN-TS](/datasets/comparison/time-series/timestral-123m-vs-tabpfn-ts-3.md): Head-to-head probabilistic loss and latency comparison.
- [Timestral-123M vs Timestral-209M](/datasets/comparison/time-series/timestral-123m-vs-timestral-209m.md): Head-to-head probabilistic loss and latency comparison.
- [Timestral-123M vs LightGBM](/datasets/comparison/time-series/timestral-123m-vs-lightgbm.md): Head-to-head probabilistic loss and latency comparison.